Files
vllm-omni/vllm_omni/deploy/mammoth_moda2_ar.yaml
T
2026-08-28 08:32:32 +08:00

15 lines
321 B
YAML

# MammothModa2 AR-only (image/video understanding and summarization).
async_chunk: false
pipeline: mammoth_moda2_ar
stages:
- stage_id: 0
devices: "0"
max_num_seqs: 16
max_model_len: 8192
gpu_memory_utilization: 0.5
enforce_eager: true
trust_remote_code: true
enable_prefix_caching: false